build: 适配沐曦

Signed-off-by: PanZezhong <[email protected]>

Author: PanZezhong1725

Cargo.lock

@@ -4,10 +4,10 @@ version = "0.0.0"
 edition.workspace = true
 
 [dependencies]
-cuda = { git = "https://github.com/YdrMaster/cuda-driver", rev = "803d64b" }
-cublas = { git = "https://github.com/YdrMaster/cuda-driver", rev = "803d64b" }
-nccl = { git = "https://github.com/YdrMaster/cuda-driver", rev = "803d64b" }
-flash-attn = { git = "https://github.com/YdrMaster/learn-flash-attn", rev = "6faa30f" }
+cuda = { git = "https://github.com/YdrMaster/cuda-driver", rev = "38f74ee" }
+cublas = { git = "https://github.com/YdrMaster/cuda-driver", rev = "38f74ee" }
+nccl = { git = "https://github.com/YdrMaster/cuda-driver", rev = "38f74ee" }
+flash-attn = { git = "https://github.com/YdrMaster/learn-flash-attn", rev = "7f96e6f" }
 nn = { git = "https://github.com/YdrMaster/InfiniNN", rev = "171c5b0" }
 ggus = { git = "https://github.com/InfiniTensor/gguf", rev = "23c362f" }
 tokeneer = { git = "https://github.com/InfiniTensor/tokeneer", rev = "c48f39f" }

llama.cu/Cargo.toml

@@ -35,6 +35,6 @@ impl Operator for AllReduce {
             .comm
             .as_ref()
             .unwrap()
-            .all_reduce(dst, Some(src), dt, ReduceType::ncclSum, stream);
+            .all_reduce(dst, Some(src), dt, ReduceType::hcclSum, stream);
     }
 }

llama.cu/src/op/all_reduce.rs

@@ -11,7 +11,7 @@ mod bindings {
             use $crate::op::random_sample::bindings::*;
             #[allow(unused_unsafe, clippy::macro_metavars_in_unsafe)]
             let err = unsafe { $f };
-            assert_eq!(err, cudaError::cudaSuccess);
+            assert_eq!(err, cudaError_t::hcSuccess);
         }};
     }
 }